# Predicting progression to proliferative diabetic retinopathy using automated versus manual quantification of retinal haemorrhages

## TL;DR

This study compares automated and manual methods for measuring retinal haemorrhages in diabetic retinopathy and finds that automated detection can predict progression to a severe form of the disease.

## Contribution

The study introduces automated quantification of retinal haemorrhages as a predictive tool for progression to proliferative diabetic retinopathy.

## Key findings

- Automated measurements of retinal haemorrhages were significantly correlated with manual grading.
- The distance of haemorrhages from the optic nerve was a significant risk factor for progression to proliferative diabetic retinopathy.
- Automated detection of haemorrhages can predict disease progression despite detecting fewer lesions than manual grading.

## Abstract

To compare automated and manual quantification of retinal haemorrhages in eyes with diabetic retinopathy (DR) and to analyse the risk of progression to proliferative DR (PDR).

Retinal haemorrhages on ultra-widefield (UWF) pseudocolor images in eyes with non-proliferative diabetic retinopathy (NPDR) were manually segmented. DR severity was assessed within the seven ETDRS fields at baseline and 1-year follow-up. Lesions were also automatically segmented using EyeRead UWF software (Eyenuk) and the frequency and area of retinal haemorrhages and the average distance of haemorrhages from the optic nerve centre were computed. Manual and automated results were compared and correlated with progression to PDR at one year.

Sixty-three eyes with NPDR at baseline were included, of which 29 progressed to PDR over one year. The automated measurements of total haemorrhage frequency, area and the distance from the optic nerve were significantly lower compared to manual grading, but the parameters were significantly correlated (r = 0.5–0.96; all P < 0.001). The distance of the haemorrhages from the optic nerve was found to be a significant risk factor for progression to proliferative DR from both approaches.

Automated detection of retinal haemorrhages may serve as a surrogate to the manual grading in predicting the progression to PDR. Although the automated approach detects a lower number of lesions compared to manual grading, the results are correlated and the automatically determined parameters are still predictive of progression.